Low Power Adaptive FIR Filter Based on Distributed Arithmetic

S Ramanathan et al. Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 5, (Part - 7) May 2016, pp.47-51
www.ijera.com 47 | P a g e
Low Power Adaptive FIR Filter Based on Distributed Arithmetic
S Ramanathan1
, Gorty Anand1
, Prasanth Reddy1
, Prof. Sri Adibhatla Sridevi2
1
( M.Tech, VLSI Design, School of Electronics Engineering (SENSE), VIT University, Vellore)
2
(Associate Professor, Micro & Nanoelectronics, School of Electronics Engineering (SENSE), VIT University,
Vellore)
ABSTRACT
This paper aims at implementation of a low power adaptive FIR filter based on distributed arithmetic (DA) with
low power, high throughput, and low area. Least Mean Square (LMS) Algorithm is used to update the weight
and decrease the mean square error between the current filter output and the desired response. The pipelined
Distributed Arithmetic table reduces switching activity and hence it reduces power. The power consumption is
reduced by keeping bit-clock used in carry-save accumulation much faster than clock of rest of the operations.
We have implemented it in Quartus II and found that there is a reduction in the total power and the core dynamic
power by 31.31% and 100.24% respectively when compared with the architecture without DA table.
Keywords: Adaptive filter, Distributed Arithmetic, Finite Impulse Response, Least Mean Square algorithm,
Lookup Table
I. INTRODUCTION
An adaptive filter tries to model the
relationship between two real time signals using an
iterative approach [1].
Four aspects defines an adaptive filter:
1) The input signals of the filter
2) The impulse response of the filter
3) The filter coefficients
4) The adaptive algorithm that is used to adjust the
weights
The LMS algorithm given by Widrow-Hoff
is used to update the tapped-delay line FIR filter's
weights because of its simplicity and convergence
performance provided by it [2].
DA based technique without multipliers [3],
are used as they provide high-throughput processing
capability and regularity that results in computing
structures that are cost-effective and area-time
efficient [4].
II. THE ADAPTIVE FILTER
In the block diagram shown in Fig.1 a
sample of x(n), the input signal, is fed to adaptive
filter and the filter outputs y(n). This output is then
compared with d(n), the desired response, and the
difference of the two is the error signal e(n) as
shown in (1).
e (n) = d (n) - y (n) (1)
Fig. 1. The general adaptive filter
Fig. 2. System identification using adaptive filter
The error signal is fed to the adaptive filter
which, in a well-defined manner, updates the
coefficients of the filter from time n to time (n + 1).
Through this adaptation process, as n increases, the
magnitude of e (n) decreases, or in other words, the
output of the adaptive filter tends to the desired
response signal.
Let x (n) be the input to an unknown
system and let dˆ (n) be its corresponding output.
Then, the desired response signal is given by
d (n) = dˆ (n) + η (n) (2)
The role of adaptive filter here is to
precisely represent dˆ (n) at its output. It can be
said that the adaptive filter driven by x (n), has
accurately modeled the unknown system if y (n) = dˆ
(n).
III. LMS ADAPTIVE ALGORITHM
In every cycle, the LMS algorithm
calculates an output and the corresponding error
value. It then uses the estimated error to update the
weights in every cycle. The LMS adaptive filter
weights in nth iteration are updated as per the
following equations.
w(n + 1) = w(n) + μ.e(n).x(n) (3a)
RESEARCH ARTICLE OPEN ACCESS

Where
e(n) = d(n) - y(n) (3b)
y(n) = wT
(n).x(n) (3c)
At the nth iteration, x(n) the input vector
and w(n) the weight vector, are respectively given
by,
x(n) = [x(n), x(n - 1),…,x(n - N + 1)]T
(4a)
w(n) = [w0(n),w1(n),…,wN-1(n)]T
(4b)
In the above equations, d(n) denotes the
desired response, y(n) denotes the filter's output in
the nth iteration, e(n) is the error computed in the nth
iteration and it is used for updating the weights, μ is
the convergence-factor and N is the length of the
filter.
In pipelined designs, only after certain
number of cycles, the feedback error e(n) becomes
available, this deley is known as the ―adaptation-
delay‖. Hence the pipelined architectures use e(n -
m) which is the delayed error in place of recent-most
error to update the current weight. Here m denotes
the adaptation-delay. The following equation is the
weight-update equation of such delayed LMS
adaptive filter
w(n + 1) = w(n) +μ.e(n - m).x(n - m) (5)
IV. DA-BASED APPROACH FOR
INNER-PRODUCT COMPUTATION
In every cycle, the LMS adaptive filter has
to compute an inner-product and this contributes to
the most of the critical-path. Let the inner-product of
(3c) be given by (6) for simplicity of presentation,
and is in fact the arithmetic sum of products which
defines the response of linear time-invariant (LTI)
network.
N-1
y = ∑ wk.xk (6)
k = 0
Where xk and wk form the N-point vectors
corresponding the recent-most N-1 input and current
weights respectively for 0 ≤ k ≤ N-1. Assuming the
bit-width of the weight to be L, each component of
weight vector can be expressed in 2’s complement
representation as shown in (7).
L-1
wk = - wk0 + ∑ wkl.2-l
(7)
l = 1
where wkl is the lth bit of wk. Substituting (7) in
(6), we get
N-1 L-1
y = ∑ xk [- wk0 + ∑ wkl.2-l
] (8)
k = 0 l = 1
Fig. 3. Conventional DA-based implementation of 4-
point inner-product.
rewriting (8), we get
N N L-1
y = - ∑ wk0.xk + ∑ ∑ wkl.xk.2-l
(9)
k = 1 k = 1 l = 1
expanding (9), we get (10)
y = - [w10.x1 + w20.x2 + w30.x3 + … + wk0.xk]
+ [w11.x1 + w21.x2 + w31.x3 + … + wk1.xk] 2-1
+ [w12.x1 + w22.x2 + w32.x3 + … + wk2.xk] 2-2
:
+ [w1(B-2).x1 + w2(B-2).x2 + … + wk(B-2).xk] 2-(B-2)
+ [w1(B-1).x1 + … +wk(B-1).xk] 2-(B-1)
(10)
Every term inside the brackets is actually a
binary AND operation that involves all the bits of
the constant and one bit of the input and the plus
signs are arithmetic sum operations. An LUT can
now be constructed which can be addressed by the
same scaled bit of all the input variables and can
access the sum of the terms inside each pair of
brackets. Fig.3 shows the LUT.
Converting the sum-of-products form of
(10) into a distributed form, we get (11)
Fig. 4. Carry-save implementation of the shift-
accumulation.

Fig. 5. DA-table for generation of the possible
sums of the input samples.
N-1 N-1 L-1
y = - ∑ wk0.xk + ∑ 2-l
. [ ∑ xk.wkl ] (11)
k = 0 k = 0 l = 1
the inner-product given by (11) is computed as
L-1 N-1
y = [ ∑ 2-l
.yl ] - y0 where yl = ∑ xk.wkl (12)
l = 1 k = 1
The partial-sum yl for l = 0, 1,2,…,L-1, can
have 2N
possible values, as the elements of the N-
point bit-sequence {wkl for 0 ≤ k ≤ N - 1} can be
either 0 or 1.
Fig. 6. The structure of DA-based LMS adaptive
filter of length N = 4.
Now, using the bit-sequence {wkl} as
address bits for the computation of inner-product, we
can read out the partial sums yl from the LUT, if we
precompute all the 2N
possible values of yl and store
it in the LUT.
Therefore, inner-product of (12) can be
calculated be calculated in L cycles of shift-
accumulation which is followed by the LUT read
operations corresponding to L number of bit-slices
{wkl} for 0 ≤ l ≤ L-1. This is shown in Fig.3. The
shift-accumulation is performed using carry-save
accumulator, as the shift-accumulation shown in
Fig.3 involves significant critical-path. This is
shown in Fig.4. The bit-slices of vector w are given
to the carry-save accumulator one after the other in
the order LSB to MSB. But in case of MSB slices,
the negative or the 2’s complement of the output of
the LUT must to be accumulated. This could easily
be achieved using XOR gates. Therefore, all the bits
of LUT output are fed to XOR gates with sign-
control input. If MSB slice appears as address, then
alone the sign-control is set to 1. So the XOR gates
produce the 1’s complement of the LUT output if
MSB slice appears as address and does not affect the
output for other cases. Lastly, the sum and carry
words that are obtained after L clock cycles are
added by an adder and the input carry of this adder is
set to 1 to account for the 2’s complement operation
of the LUT output for the MSB slice.
Fig. 7. (a) Structure of 4-point inner-product block.
(b) Structure of weight increment block for N = 4.
(c) The logic which is used for the generation of the
control word t for the barrel-shifter for L = 8.
Fig. 8. Structure of DA-based LMS AF with N =
16 and P = 4.
The content of kth LUT location is expressed as
N-1
ck = ∑ xj.kj (13)
j = 0

where kj is (j + 1)th bit of the N-bit binary
representation of integer k where k lies in the range
0 ≤ k ≤ 2N
- 1. We can pre-compute ck for 0 ≤ k ≤ 2N
- 1 and store it in a RAM-based LUT of 2N
words.
But here we store (2N
- 1) words in a DA-table that
consists of 2N
- 1 registers, in place of storing 2N
words in LUT. Fig.5 shows an example of one such
DA-table for N = 4. It has only 15 registers that are
used to store the sums of input words which are pre-
computed. Seven adders in parallel, compute seven
new values of ck.
V. DA BASED ADAPTIVE FILTER
STRUCTURE
Fig.6 shows the structure of DA-based adaptive filter
of filter length N = 4. It has a 4-point inner-product
and a block for weight-increment, along with the
additional circuits required for the computation of
e(n), the error value and the control word t for
barrel-shifters.
As shown in Fig.7a, the 4-point inner-product block
consists of a DA-table that has an array of 15
registers that store the partial-inner-products yl for l
in range 0 < l ≤ 15 and to select the content of one of
these registers, a 16 : 1 MUX which is used.
Bit-slices of weights A = {w3l w2l w1l w0l}
for 0 ≤ l ≤ L - 1 are given to the MUX as control in
the order LSB to MSB, and the output of MUX is
given to carry-save accumulator which is shown in
Fig. 4.
The carry-save accumulator shift-
accumulates all the partial-inner-products after L bit-
cycles, and generates two words of size (L + 2)-bit
each during these L bit-cycles, one for sum and other
one for the carry. The sum and carry words are
shifted-added with an input carry ’1’ in order to
generate filter output which is subtracted from d(n),
the desired output, later to obtain the error value
e(n).
By assuming N = PQ, we can now
decompose inner-product computation of (6) into N
= P small adaptive filtering blocks of length P as
shown in (14)
P-1 2P-1 N-1
y = ∑ wk.xk + ∑ wk.xk + … + ∑ wk.xk
(14)
k = 0 k = P k = N-P
Each of the above mentioned P-point inner-
product computation blocks, to update P weights,
have weight-increment unit. The structure for N = 16
and P = 4 is shown in the Fig.8. It has four inner-
product blocks of length P = 4 as shown in Fig.7a.
Two separate binary adder-trees are used to add the
(L + 2)-bit carry and sum produced by these four
blocks. Four carry-in bits are added to the sum
words which are the output of four 4-point inner-
product blocks. At the first level binary adder-tree of
the carry words, two carry-in bits are set as the input
carry, as carry words are of twice the weight as
compared to sum words. This is same as inclusion of
four carry-in to sum words.
To make the length of sign-magnitude
separator as L-bit, assuming μ = 1/N, we can
truncate the 4 LSBs of error e(n) for N = 16. The
performance of adaptive filter is not very much
affected by the truncation as the design
requires the location of most significant 1 of μ.e(n).
VI. RESULTS
The Low Power Adaptive FIR Filter Based
on Distributed Arithmetic has been implemented in

Quartus II and we have compared it with an
architecture without distributed arithmetic table.
TABLE I. POWER CONSUMPTION
Architecture Power Consumption
Total Power Core Dynamic Power
Conventional
FIR Filter
110.77mW 24.91mW
DA-Based
LMS Adaptive
FIR Filter
84.36mW 12.44mW
It is found that there is a reduction in the
total power and the core dynamic power by 31.31%
and 100.24% respectively when compared with the
architecture without DA table.
VII. CONCLUSION
We have implemented an efficient
pipelined architecture for high-throughput, low-
power and low-area DA-based adaptive filter.
Throughput rate is quite significantly improved by
means of concurrent processing of weight-update
and filtering operation and the parallel LUT update.
For computation of the filter output, we have used a
carry-save accumulation unit for signed partial-
inner-products computation.
ACKNOWLEDGEMENTS
We have implemented a Low Power
Adaptive FIR Filter Based on Distributed Arithmetic
under the supervision of Prof. Sri Adibhatla Sridevi,
Associate Professor, SENSE. We thank her for
providing proper guidance and timely help which
helped us to successfully complete the work.
REFERENCES
Books:
[1]. Douglas S.C. Introduction to Adaptive
Filters, Digital Signal Processing
Handbook, Ed. Vijay K. Madisetti and
Douglas B. Williams, Boca Raton: CRC
Press LLC, 1999.
[2]. S. Haykin and B. Widrow, Least-mean-
square adaptive filters. Wiley-Interscience,
Hoboken, NJ, 2003.
Journal Papers:
[3]. S. A. White, ―Applications of the
distributed arithmetic to digital signal
processing: A tutorial review,‖ IEEE ASSP
Magazine, vol. 6, no. 3, pp. 5–19, Jul. 1989.
[4]. S. Y. Park And P. K. Meher, "Low-Power,
High-Throughput, and Low-Area Adaptive
FIR Filter Based On DA" IEEE
Transactions on Circuits And Systems—II,
Express Briefs, vol. 60, no. 6, June 2013,
pp. 346-350.

Low Power Adaptive FIR Filter Based on Distributed Arithmetic

Recommended

More Related Content

What's hot (17)

Similar to Low Power Adaptive FIR Filter Based on Distributed Arithmetic (20)

Recently uploaded (20)

Low Power Adaptive FIR Filter Based on Distributed Arithmetic